HAL Id: hal-00759119
https://hal.archives-ouvertes.fr/hal-00759119
Preprint submitted on 30 Nov 2012
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires
Local feedback stabilisation to a non-stationary solution for a damped non-linear wave equation
Kaïs Ammari, Thomas Duyckaerts, Armen Shirikyan
To cite this version:
Kaïs Ammari, Thomas Duyckaerts, Armen Shirikyan. Local feedback stabilisation to a non-stationary
solution for a damped non-linear wave equation. 2012. �hal-00759119�
Local feedback stabilisation to a non-stationary solution for a damped non-linear wave equation
Ka¨ıs Ammari
∗Thomas Duyckaerts
†Armen Shirikyan
‡Abstract
We study a damped semi-linear wave equation in a bounded domain of R
3with smooth boundary. It is proved that any H
2-smooth solution can be stabilised locally by a finite-dimensional feedback control supported by a given open subset satisfying a geometric condition. The proof is based on an investigation of the linearised equation, for which we construct a stabilising control satisfying the required properties. We next prove that the same control stabilises locally the non-linear problem.
AMS subject classifications: 35L71, 93B52, 93B07
Keywords: Non-linear wave equation, distributed control, feedback sta- bilisation, truncated observability inequality
Contents
0 Introduction 2
1 Preliminaries 5
1.1 Cauchy problem for a semi-linear wave equation . . . . 5
1.2 Cauchy problem for the wave equation with low-regularity data . 6 1.3 Commutator estimates . . . . 9
2 Stabilisation of the linearised equation 12 2.1 Main result and scheme of its proof . . . . 12
2.2 Construction of a stabilising control . . . . 14
2.3 Dynamic programming principle and feedback law . . . . 17
2.4 Conclusion of the proof of Theorem 2.3 . . . . 19
∗D´epartement de Math´ematiques, Facult´e des Sciences de Monastir, Universit´e de Mona- stir, 5019 Monastir, Tunisie; E-mail: Kais.Ammari@fsm.rnu.tn
†D´epartement de Math´ematiques, Institut Galil´ee, Universit´e Paris 13, 99 avenue Jean- Baptiste Cl´ement, 93430 Villetaneuse, France; E-mail: duyckaer@math.univ-paris13.fr
‡Department of Mathematics, University of Cergy–Pontoise, CNRS UMR 8088, 2 avenue Adolphe Chauvin, 95302 Cergy–Pontoise, France; E-mail: Armen.Shirikyan@u-cergy.fr
3 Observability inequalities 20 3.1 Observability of low-regularity solutions . . . . 20 3.2 Truncated observability inequality . . . . 23 3.3 Observability inequality for the wave equation . . . . 24 4 Main result: stabilisation of the non-linear problem 25
Bibliography 27
0 Introduction
Let us consider the damped non-linear wave equation (NLW)
∂
t2u + γ∂
tu − ∆u + f (u) = h(t, x), x ∈ Ω, (0.1) where Ω ⊂ R
3is a bounded domain with a smooth boundary ∂Ω, γ > 0 is a parameter, h is a locally square-integrable function with range in L
2(Ω), and f : R → R is a function satisfying some natural growth and regularity conditions ensuring the existence, uniqueness, and regularity of a solution. Equation (0.1) is supplemented with the Dirichlet boundary condition,
u
∂Ω= 0, (0.2)
and the initial conditions
u(0, x) = u
0(x), ∂
tu(0, x) = u
1(x), (0.3) where u
0∈ H
01(Ω) and u
1∈ L
2(Ω). It is well known that, even though prob- lem (0.1), (0.2) is dissipative and possesses a global attractor (which is finite- dimensional in the autonomous case), its flow is not locally stable, unless we impose very restrictive conditions on the nonlinear term f . That is, the differ- ence between two solutions with close initial data, in general, grows in time. The purpose of this paper is to show that any sufficiently regular solution of (0.1), (0.2) can be stabilised with the help of a finite-dimensional feedback control localised in space. Namely, instead of (0.1), consider the controlled equation
∂
2tu + γ∂
tu − ∆u + f (u) = h(t, x) + η(t, x), x ∈ Ω, (0.4) where η is a control whose support in x contains a subset satisfying a geometric condition. For a time-dependent function v, we set
Φ
v(t) =
v(t), v(t) ˙
, E
v(t) = Z
Ω
| v(t, x)| ˙
2+ |∇
xv(t, x)|
2dx,
where the dot over a function stands for its time derivative. The following
theorem is an informal statement of the main result of this paper.
Main Theorem. Let u ˆ be the solution of (0.1)–(0.3) with sufficiently regular initial data and let ω ⊂ Ω be a neighbourhood of the boundary ∂Ω. Then there is a finite-dimensional subspace F ⊂ H
01(ω) and a family of continuous operators K
uˆ(t) : H
01× L
2→ F such that, for any [u
0, u
1] ∈ H
01× L
2that is sufficiently close to Φ
uˆ(0), problem (0.2)–(0.4) with η(t) = K
uˆ(t)Φ
u−uˆ(t) has a unique solution u(t, x), which satisfies the inequality
E
u−uˆ(t) ≤ Ce
−βtE
u−uˆ(0), t ≥ 0, (0.5) where C and β are positive numbers not depending on [u
0, u
1].
We refer the reader to Section 4 (see Theorem 4.1) for the exact formulation of our result. Before outlining the main idea of the proof of this theorem, we discuss some earlier results concerning the semi-linear wave equation with localised control. This problem was studied in a number of works, and first local results were obtained by Fattorini [Fat75] for rectangular domains and Chewning [Che76] for a bounded interval. Zuazua [Zua90a, Zua93] proved that the 1D wave equation with a nonlinear term f (u) growing at infinity no faster that u(log u)
2possesses the property of global exact controllability, provided that the control time is sufficiently large. These results were extended later to the multidimensional case, as well as to the case of nonlinearities with a faster growth and “right” sign at infinity [Zha00, LZ00]. Furthermore, the question of stabilisation to the zero solution was studied in [Zua90b].
The application of methods of microlocal analysis enables one to get sharper results. In the linear case, the exact controllability and stabilisation to a station- ary solution are established under conditions that are close to being necessary;
e.g., see [BLR92, LR97]. Dehman, Lebeau, and Zuazua [DLZ03] proved the global exact controllability by a control supported in the neighbourhood of the boundary, provided that nonlinearity is subcritical and satisfies the inequali- ties f (u)u ≥ 0; see also [DL09] for a refinement of this result. Fu, Yong, and Zhang [FYZ07] established a similar result for the equations with variable coef- ficients and a nonlinearity f growing at infinity slower than u(log u)
1/2. Coron and Tr´elat [CT06] proved exact controllability of 1D nonlinear wave equation in a connected component of steady states. Laurent [Lau11] extended the result of [DLZ03] to the case of the nonlinear Klein–Gordon equation with a critical exponent. Very recently, Joly and Laurent [JL12] proved the global exact con- trollability of the NLW equation of a particular form, using a fine analysis of the dynamics on the attractor. In conclusion, let us mention that the problem of exact controllability and stabilisation for other type of semi-linear dispersive equations was investigated in a large number of works, and we refer the reader to the papers [DGL06, Lau10] for an overview of the literature in this direction.
To the best of our knowledge, the problem of stabilisation of a nonstation- ary solution ˆ u by a finite-dimensional localised control was not studied earlier.
Without going into detail, let us describe informally the main idea of our ap- proach, which is based on the study of Eq. (0.1) linearised around ˆ u. We thus consider the equation
∂
t2u + γ∂
tu − ∆u + b(t, x)u = η(t, x), x ∈ Ω, (0.6)
supplemented with the initial and boundary conditions (0.2) and (0.3). We wish to find a finite-dimensional localised control depending on the initial conditions [u
0, u
1] such that the energy E
u(t) goes to zero exponentially fast. Following a well-known idea coming from the theory of attractors (see [Har85] and Sec- tion II.6 in [BV92]), we represent a solution of (0.6), (0.3) in the form u = v +w, where v is the solution of (0.6) with b ≡ η = 0 issued from [u
0, u
1]. Then w satisfies the zero initial conditions and Eq. (0.6) with η replaced by η − bv. Let us note that v goes to zero in the energy space exponentially fast and that w has better regularity properties than v. It follows, in particular, that the decay of a sufficiently large finite-dimensional projection of w will result in exponential stabilisation of u. Combining this with the general scheme used in [BRS11] and a new observability inequality established in Section 3, we construct a finite- dimensional localised control η which squeezes to zero the energy norm of w and that of u. A standard technique enables one to prove the latter property can be achieved by a feedback control. We refer the reader to Section 2 for more details.
The paper is organised as follows. In Section 1, we recall some well-known results on the Cauchy problem for semi-linear wave equation and establish the existence and uniqueness of a solution for the linear problem with low-regularity data. Section 2 is devoted to the stabilisation to zero for the linearised equation by a finite-dimensional localised control. The key tool for proving this result is the truncated observability inequality established in Section 3. Finally, the main result on local stabilisation for the non-linear problem is presented in Section 4.
Acknowledgements . This work was initiated when the third author was visiting The University of Monastir (Tunisia) in April of 2011. He thanks the institute for hospitality. The research of TD was supported by the ERC starting grant DISPEQ , the ERC advanced grant BLOWDISOL (No. 291214), and the ANR JCJC grant SchEq. The research of AS supported by the Royal Society–
CNRS grant Long time behavior of solutions for stochastic Navier–Stokes equa- tions (No. YFDRN93583) and the ANR grant STOSYMAP (No. ANR 2011 BS01 015 01).
Notation
Let J ⊂ R be a closed interval, let Ω ⊂ R
dbe a bounded domain with C
∞boundary ∂Ω, and let X and Y be Banach spaces. We shall use the following functional spaces.
C
0∞(Ω) is the space of infinitely smooth functions f : Ω → R with compact support.
L
p= L
p(Ω) is the usual Lebesgue space of measurable functions f : Ω → R such that
kf k
pLp:=
Z
Ω
|f (x)|
pdx < ∞.
When p = 2, this norm is generated by the L
2-scalar product (·, ·) and is denoted
by k · k.
H
s= H
s(Ω) is the Sobolev space of order s with the standard norm k · k
s. H
0s= H
0s(Ω) is the closure of C
0∞(Ω) in H
s. It is well known that H
0s= H
sfor s <
12and that H
swith s ≤ 0 is the dual of H
0−swith respect to (·, ·).
C(J, X ) is the space of continuous functions f : J → X with the topology of uniform convergence on bounded intervals. Similarly, C
k(J, X) is the space of k time continuously differentiable functions f ∈ C(J, X ).
L
p(J, X) is the space of Bochner measurable functions f : J → X with a finite norm
kf k
Lp(J,X)= Z
J
kf (t)k
pXdt
1/p. In the case p = ∞, this norm should be replaced by ess sup
t∈J
kf (t)k
X.
W
k,p(J, X) is the space of functions f ∈ L
p(J, X) such that ∂
tjf ∈ L
p(J, X ) for j = 0, . . . , k.
L(X, Y ) denotes the space of continuous linear operators from X to Y with the usual operator norm k ·k
L(X,Y). When the choice of X and Y is clear, we simply write k · k
L.
Given a function of the time variable v(t), we write Φ
v(t) = [v(t), v(t)], where ˙ the dot stands for the time derivative.
We denote by C unessential numbers (which may vary from line to line) and by M
i(a
1, . . . , a
n) positive constants depending on the parameters a
1, . . . , a
n. We write J
T= [0, T ], J
T(s) = [s, s + T ], R
+= [0, +∞), and R
s= [s, +∞).
1 Preliminaries
In this section, we first recall a well-known result on the existence, uniqueness, and regularity of a solution for (0.1)–(0.3). We next turn to the linearised problem, for which we prove the well-posedness in a space of functions of low regularity. Finally, we derive some commutator estimates used in what follows.
1.1 Cauchy problem for a semi-linear wave equation
Let us consider the equation
∂
t2u + γ∂
tu − ∆u + f (u) = g(t, x), x ∈ Ω, (1.1) where Ω ⊂ R
3is a bounded domain with the boundary ∂Ω ∈ C
∞, γ > 0 is a parameter, g is a locally square-integrable function of time with range in L
2(Ω), and f ∈ C
1(R) is a function vanishing at u = 0 and satisfying the inequalities
|f
′(u)| ≤ C(1 + |u|
2), F(u) :=
Z
u 0f (v) dv ≥ −C, u ∈ R. (1.2)
Equation (1.1) is supplemented with the initial and boundary conditions (0.2) and (0.3). A proof of the following result on the well-posedness of the initial- boundary value problem for (1.1) and regularity of solutions can be found in Chapter 1 of [Lio69] and Section 1.8 of [BV92].
Proposition 1.1. Under the above-mentioned hypotheses, for any u
0∈ H
01(Ω), u
1∈ L
2(Ω), and g ∈ L
2loc(R
+, L
2), problem (1.1), (0.2), (0.3) has a unique solution
u ∈ C(R
+, H
01) ∩ C
1(R
+, L
2) ∩ W
loc2,2(R
+, H
−1). (1.3) Moreover, if in addition u
0∈ H
2(Ω), u
1∈ H
01(Ω), and ∂
tg ∈ L
2loc(R
+, L
2), then u ∈ C(R
+, H
2) ∩ C
1(R
+, H
01) ∩ W
loc2,2(R
+, L
2). (1.4) We now formulate a result on the time boundedness of solution for (1.1) under some additional assumptions on f . Namely, let us assume that f ∈ C
2(R) is such that
f (u)u ≥ c F (u) − C, f
′(u) ≥ −C, |f
′′(u)| ≤ C(1 + |u|), u ∈ R, (1.5) where C and c are positive constants. Note that these conditions are satisfied for polynomials of degree 3 with positive leading coefficient and, more generally, for C
2-smooth functions behaving at infinity as c|u|
ρ−1u with c > 0 and ρ ∈ [1, 3].
The following result is established by Zelik [Zel04].
Proposition 1.2. Let us assume that f ∈ C
2(R) satisfies (1.2) and (1.5) and let g ∈ W
1,∞(R
+, L
2). Then, for any u
0∈ H
2(Ω) ∩ H
01(Ω) and u
1∈ H
01(Ω), the solution u(t, x) of (1.1), (0.2), (0.3) satisfies the inequality
ku(t)k
2+ k u(t)k ˙
1≤ M
1ku
0k
2, ku
1k
1, G
, t ≥ 0, (1.6)
where we set
G := ess sup
t≥0
kg(t, ·)k + k∂
tg(t, ·)k
< ∞.
1.2 Cauchy problem for the wave equation with low-regu- larity data
In this and the next subsections, we assume that the space dimension d ≥ 1 is arbitrary, even though the results obtained here will be used only for d = 3. We study the linearised problem
¨
v + γ v ˙ − ∆v + b(t, x)v = η(t, x), x ∈ Ω, (1.7) v
∂Ω= 0, (1.8)
v(0, x) = v
0(x), v(0, x) = ˙ v
1(x), (1.9)
where b and η are functions of low regularity, and γ ∈ R. Namely, let {e
j} be
the complete set of L
2-normalised eigenvectors for the Dirichlet Laplacian in Ω
(denoted by −∆) and let {λ
j} be the corresponding eigenvalues numbered in an increasing order. We define the scale of spaces associated with −∆ by the relation
H
Ds= H
Ds(Ω) = n
f ∈ L
2(Ω) : X
j≥1
(f, e
j)
2λ
sj< ∞ o
for s ≥ 0 (1.10) and denote by H
D−s= H
D−s(Ω) the dual of H
Dswith respect to the L
2scalar product. It is well known that (see [Fuj67])
H
Ds= H
0sfor
12< s <
12, H
Ds= H
sfor −
32< s <
12. (1.11) The wave propagator for (1.7), (1.8) is well defined (with the help of the eigenfunction expansion) when b ≡ η ≡ 0. In this case, for any initial data [v
0, v
1] ∈ H
Ds:= H
Ds× H
Ds−1there is a unique solution
v ∈ C(R, H
Ds) ∩ C
1(R, H
Ds−1), which satisfies the inequality
kΦ
v(t)k
HsD≤ e
c|t|kΦ
v(0)k
HsDfor t ∈ R, (1.12) where c ≥ 0 depends only on γ. We denote by S(t) : H
Ds→ H
sDthe oper- ator taking [v
0, v
1] to Φ
v(t) and write S(t) = [S
0(t), S
1(t)], so that S
0(t) is a continuous operator from H
sDto H
Dsfor any s ∈ R.
Let us define H
s= H
0s× H
0s−1and note that, in view of (1.11), we have H
s= H
sDfor s ∈ (−
12,
12). The proof of the following result is rather standard and is based on the Duhamel representation, the Banach fixed point theorem, and an estimate for the Sobolev of the product of two functions.
Proposition 1.3. Let r and T be positive numbers, let J
T= [0, T ], and let b ∈ L
∞(J
T× Ω) ∩ L
∞(J
T, H
r). Then there is σ
0(r) > 0 such that for any σ ∈ [0, σ
0(r)], [v
0, v
1] ∈ H
−σ, and η ∈ L
1(J
T, H
−σ−1) problem (1.7)–(1.9) has a unique solution
v ∈ X
Tσ:= C(J
T, H
−σ) ∩ C
1(J
T, H
−σ−1).
Moreover, there is M
2= M
2T, kbk
L∞(JT,L∞∩Hr)such that
sup
t∈JT
kΦ
v(t)k
H−σ≤ M
2[v
0, v
1]
H−σ+ η
L1(JT,H−σ−1). (1.13)
Note that inequality (1.13) is true for σ = −1, provided that b ∈ L
∞(J
T×Ω) for any T > 0; this is a simple consequence of the standard energy estimate for the wave equation and the Gronwall inequality.
Proof of Proposition 1.3. We need to construct a solution of the integral equa- tion
v(t) = S
0(t)[v
0, v
1]−
Z
t 0S
0(t−s)
0, b(s)v(s) ds+
Z
t 0S
0(t−s) 0, η(s)
ds, (1.14)
where t ∈ J
T. We shall prove the existence of a solution on small time interval J
τ= [0, τ ] whose length does not depend on the size of the initial data and the right-hand side. The global existence will then follow by iteration.
Step 1. Bound on the Duhamel term. Let us set
Qg(t) = Z
t0
S
0(t − s)[0, g(s)] ds.
We first show that if g ∈ L
1(J
τ, H
−1−σ), then
Qg ∈ X
τσ, kQgk
Xτσ≤ C
1kgk
L1(Jτ,H−1−σ), (1.15) where C
1does not depend on g. Indeed, if 0 ≤ σ <
12, then H
−σ−1= H
D−σ−1and thus the mapping s 7→ {S
0(t − s)[0, g(s)], t ∈ J
τ} belongs to the space L
1(J
τ, Y
τσ), where Y
τσ= C(J
τ, H
−σ). Furthermore, by (1.12), we have
kS
0(t − s)[0, g(s)]k
−σ≤ e
cτkg(s)k
−σ−1for all t, s ∈ J
τ, It follows that Qg ∈ C(J
τ, H
−σ) and
kQg(t)k
−σ≤ C kgk
L1(Jτ,H−σ−1)for t ∈ J
τ. Using the relation
∂
∂t S
0(t − s)[0, g(s)] = S
0(t − s)[g(s), 0], we see that ∂
t(Qg) ∈ C(J
τ, H
−σ−1) and
k∂
tQg(t)k
−σ−1≤ C kgk
L1(Jτ,H−σ−1), t ∈ J
τ. This completes the proof of (1.15).
Step 2. Fixed point argument. We shall need the following lemma, whose proof is given at the end of this subsection.
Lemma 1.4. Let r ∈ [0, 1 + d/2], let Ω ⊂ R
dbe a bounded domain with smooth boundary, and let a ∈ L
∞∩ H
r. Then, for any s ∈ [0, r], we have
kaf k
s≤ C kak
L∞∩Hrkf k
psfor f ∈ H
ps, (1.16) kaf k
−sp≤ C kak
L∞∩Hrkf k
−sfor f ∈ H
−s, (1.17) where p =
d+22r, and C > 0 depends only on Ω.
For v ∈ X
τσ, we define P v(t) =
Z
t 0S
0(t − s)
0, b(s)v(s)
ds, t ∈ J
τ. (1.18) Using (1.17), one can find a number σ
0(r) > 0 such that, if 0 < σ ≤ σ
0(r), then bv ∈ L
∞(J
τ, H
−σ−1) and
kbvk
L∞(Jτ,H−σ−1)≤ C kbk
L∞(Jτ,L∞∩Hr)kvk
Xτσ.
Hence, in view of (1.15), we have
kP vk
Xτσ≤ C τ kbk
L∞(Jτ,Hr∩L∞)kvk
Xτσ. (1.19) Now note that (1.14) holds for all t ∈ J
τand a function v ∈ X
τσif and only if
(Id + P )u = S(t)[v
0, v
1] + Qη, u ∈ X
τσ. Choosing τ so small that
Cτ kbk
L∞(0,τ;Hr∩L∞)≤ 1 2 , we see from (1.19) that Id + P is invertible in X
τσand
(Id + P )
−1 L(Xτσ,Xτσ)≤ 2. (1.20) We thus obtain the existence of a unique solution v ∈ X
τσfor (1.14), which can be represented in the form
v = (Id + P)
−1(S(·)[v
0, v
1] + Qη) .
Combining this with (1.12), (1.15), and (1.20), we obtain the required esti- mate (1.13). This completes the proof of the proposition.
Proof of Lemma 1.4. Let us consider the multiplication operator f 7→ af . Using the continuity of the embedding H
1+d/2⊂ L
∞, the fact that H
sis a Banach algebra for s > d/2, and interpolation techniques, it is easy to prove that
kaf k
r≤ C
1kak
rkf k
1+d/2, f ∈ H
1+d/2. On the other hand, it is obvious that
kaf k ≤ kak
L∞kf k, f ∈ L
2. By interpolation, the above two inequalities imply that
kaf k
θr≤ C
1θkak
θrkak
1−θL∞kf k
θ(1+d/2), f ∈ H
θ(1+d/2). Taking θ = s/r, we arrive at inequality (1.16).
To prove (1.17), note that p ≥ 1, whence it follows that the operator of multiplication by a sends H
0spto H
0s. By duality, it is also continuous from H
−sto H
−ps, and inequality (1.17) is implied by (1.16).
1.3 Commutator estimates
Given a function ψ ∈ C
0∞(R), we define ψ(−∆)f =
X
∞j=1
ψ(λ
j)(f, e
j)e
j, f ∈ L
2.
The aim of this subsection is to derive some estimates for the commutator
of ψ(−∆) with the multiplication operator. In what follows, given a function a ∈
L
2(Ω), we denote by the same symbol the corresponding multiplication operator
sending f to af . We begin with the case of a smooth function.
Lemma 1.5. Let a ∈ C
∞(Ω), ψ ∈ C
0∞(R), and α ∈ [0, 1/2). Then there is M
3= M
3(a, ψ) such that, for h ∈ (0, 1], we have
ψ(−h
2∆), a
L(Hα+10 ,Hα)
≤ M
3h
2, (1.21)
ψ(−h
2∆), a
L(H−α,L2)≤ M
3h
1−α. (1.22) Proof. We first prove (1.21). Using the Fourier inversion formula, we get
ψ(−h
2∆) = 1 2π
Z
R
e
−ish2∆ψ(s) ˆ ds, where ˆ ψ is the Fourier transform of ψ. Thus
ψ(−h
2∆), a f = 1
2π Z
R
ψ(s)v(h ˆ
2s) ds, where v(s) =
e
−is∆, a
f is the solution of the problem
− i∂
sv + ∆v = [a, ∆](e
−is∆f ), v
s=0= 0. (1.23)
Using the fact that H
0α+1= H
Dα+1, we derive [a, ∆] e
−is∆f
α=
(∆ a)e
−is∆f + 2∇a · ∇(e
−is∆f )
α≤ C
e
−is∆f
α+1= C kf k
α+1. Combining this with (1.23), we obtain
kv(s)k
α≤ C|s|kf k
α+1for all s ≥ 0, whence it follows that
ψ(−h
2∆), a f
α≤ C Z
R
kv(h
2s)k
α| ψ(s)| ˆ ds ≤ C h
2kf k
α+1s ψ ˆ
L1. This completes the proof of (1.21).
To prove (1.22), we first note that inequality (1.21) with α = 0 implies by duality that
a, ψ(−h
2∆)
L(L2,H−1)≤ Ch
2. (1.24) Let ϕ ∈ C
0∞(R) be such that ϕ = 1 on the support of ψ. Using the relation ψ(−h
2∆)ϕ(−h
2∆) = ψ(−h
2∆), we get
a, ψ(−h
2∆)
=
a, ψ(−h
2∆)
ϕ(−h
2∆) + ψ(−h
2∆)
a, ϕ(−h
2∆) . Combining with (1.21), (1.24), and the inequality
kϕ(−h
2∆)k
L(L2,H0α+1)+ kψ(−h
2∆)k
L(H−1,Hα)≤ Ch
−1−α, we derive
a, ψ(−h
2∆)
L(L2,Hα)≤ Ch
1−α.
By duality, we obtain (1.22).
We now turn to the case of functions of low regularity, which will be impor- tant in the derivation of an observability inequality (see Section 3).
Lemma 1.6. For any r > 0 there is σ > 0 such that, if ψ ∈ C
0∞(R) and a ∈ H
r(Ω) ∩ L
∞(Ω), then
ψ(−h
2∆), a
L(H−σ,H−1)≤ M
4h
σ, (1.25) where M
4= M
4(ψ, kak
L∞∩Hr).
Proof. There is no loss of generality in assuming that 0 < r ≤ 1. Let us set A
h=
a, ψ(−h
2∆)
. Since ψ(−h
2∆) : L
2→ L
2is bounded uniformly in h ∈ (0, 1], we have
sup
h∈(0,1]
kA
hk
L(L2,L2)< ∞. (1.26) Furthermore, since ψ(−h
2∆) : H
Ds→ H
Dsis also bounded uniformly in h ∈ (0, 1]
for any s ∈ R, it follows from (1.16) that sup
h∈(0,1]
kA
hk
L(HpsD,HDs)< ∞ for s ∈ [0, r], (1.27) where p =
d+22r. We next show
sup
h∈(0,1]
h
−2kA
hk
H2D→H−2D
< ∞. (1.28) To this end, we write (cf. proof of Lemma 1.5)
A
hf = 1 2π
Z
R
ψ(τ)v(h ˆ
2τ) dτ, v(t) = ae
−it∆f − e
−it∆(af ). (1.29) Now note that
∂
tv = i∆ e
it∆(af )
− ia∆ e
it∆f
, v(0) = 0.
It follows that k∂
tvk
H−2D
≤
e
it∆(af )
L2+ C
∆e
it∆f
L2≤ Ckf k
2H2 D, whence we conclude that
kvk
H−2D
≤ C|t|kf k
HD2. Recalling (1.29), we get the inequality
kA
hf k
H−2D
≤ Ch
2τ ψ ˆ
L1kf k
H2D
, which implies (1.28).
Interpolating (1.26) and (1.28), we derive sup
h∈(0,1]
h
−skA
hk
L(HsD,HD−s)
< ∞.
Interpolating with (1.27), we deduce sup
h∈(0,1]
h
−s/3kA
hk
H(2p+1)s/3D →HDs/3
< ∞.
Taking s = 3/(2p + 1), by duality we obtain inequality (1.25) with σ = s/3.
2 Stabilisation of the linearised equation
This section is devoted to the stabilisation of the linearised problem (1.7), (1.8), in which γ > 0, b is a given function, and η is a finite-dimensional control supported by a given subdomain of Ω. The main result of this section is the existence of a feedback control exponentially stabilising problem (1.7), (1.8). To this end, we first construct a finite-dimensional stabilising control and then use a standard technique to get a feedback law.
2.1 Main result and scheme of its proof
As before, we denote by Ω ⊂ R
3a bounded domain with C
∞boundary Γ. We shall always assume that the following two conditions are satisfied.
Condition 2.1. The smooth surface Γ has only finite-order contacts with its tangent straight lines.
In other words, let y ∈ Γ and τ
y⊂ R
3be the tangent plane to Γ at the point y. In a small neighbourhood of y, the surface Γ can be represented as the graph of a smooth function ϕ
y: τ
y→ R vanishing at y together with its first-order derivatives. Condition 2.1 requires that the restriction of ϕ
yto the straight lines passing through y has no zero of infinite order at the point y.
To formulate the second condition, we first introduce some notation. Given x
0∈ R
3, define Γ(x
0) as the set of points y ∈ Γ such that hy − x
0, n
yi > 0, where n
ystands for the outward unit normal to Γ at the point y. Let ω be the support of the control function η entering (1.7).
Condition 2.2. There is x
0∈ R
3\ Ω and δ > 0 such that
Ω
δ(x
0) := {x ∈ Ω : there is y ∈ Γ(x
0) such that |x − y| < δ} ⊂ ω.
Before formulating the main result of this section, let us make some com- ments about the above hypotheses. Condition 2.2 naturally arising in the con- text of the multiplier method (see [Lio88]) ensures that the observability in- equality holds for (1.7) in the energy norm. On the other hand, Condition 2.1 enables one to define a generalised bicharacteristic flow on Ω (see Section 24.3 in [H¨or94]). Together with Condition 2.2, this implies that if T is sufficiently large, then for any δ
′∈ (0, δ) the pair (Ω
δ′(x
0), T ) geometrically controls Ω in the sense that every generalised bicharacteristic ray of length T meets the set Ω
δ′(x
0). In view of [BLR92], it follows that the observability inequality holds for Eq. (1.7) with b ≡ 0 in spaces of negative regularity. We shall combine these two results with some commutators estimates and a compactness argument to establish a truncated observability inequality for (1.7) (see Section 3.2), which is a key point of the proof of the theorem below.
Let us fix a function χ ∈ C
0∞(R
3) such that supp χ ∩ Ω ⊂ ω and χ(x) = 1 for
x ∈ Ω
δ/2(x
0). We denote by F
mthe vector span of the functions {χe
1, . . . , χe
m},
where {e
j} is a complete set of L
2normalised eigenfunctions for the Dirichlet
Laplacian. The following theorem is the main result of this section.
Theorem 2.3. Let Condition 2.1 and 2.2 be satisfied, let R and r be positive numbers, and let b(t, x) be a function such that
|b| := ess sup
t≥0
kb(t, ·)k
L∞∩Hr≤ R. (2.1) Then there is an integer m ≥ 1, positive numbers C and β, and a family of continuous linear operators
K
b(t) : H
01× L
2→ F
m, t ≥ 0, such that the following properties hold.
Time continuity and boundedness. The function t 7→ K
b(t) is continuous from R
+to the space L(H
01× L
2, F
m) endowed with the weak operator topology, and its norm is bounded by C.
Exponential decay. For any s ≥ 0 and [v
0, v
1] ∈ H
01× L
2, problem (1.7), (1.8) with the right-hand side η(t) = K
b(t)[v(t), v(t)] ˙ has a unique solution v ∈ C(R
+, H
01) ∩ C
1(R
+, L
2) satisfying the initial conditions
v(s, x) = v
0(x), ∂
tv(s, x) = v
1(x). (2.2) Moreover, we have the inequality
E
v(t) ≤ C e
−β(t−s)E
v(s), t ≥ s. (2.3) Let us sketch the proof of this result. We first prove that, for any β > 0 and sufficiently large T > 0, there is a linear operator Θ
s: H → L
2(J
T(s), F
m), where J
T(s) = [s, s +T ] and H = H
01× L
2, such that the norm of Θ
sis bounded uniformly in s ≥ 0, and for any [v
0, v
1] ∈ H
01× L
2the solution of problem (1.7), (1.8), (2.2) with η = Θ
s[v
0, v
1] satisfies the inequality
E
v(s + T ) ≤ e
−βTE
v(s). (2.4) For given initial data [v
0, v
1] ∈ H
01× L
2, an exponentially stabilising control η can be constructed by the rule
η
JT(0)= Θ
0[v
0, v
1], η
JT(kT)= Θ
kTΦ
v(kT ), k ≥ 1. (2.5) Inequality (2.4) and the uniform boundedness of Θ
simply that (2.3) holds with s = 0. Once the existence of at least one exponentially stabilising control is proved, one can use a standard technique based on the dynamical programming principle to construct a feedback law possessing the required properties. The uniqueness of a solution is proved by a standard argument based on the Gronwall inequality.
The rest of this section is organised as follows. In Subsection 2.2, we prove
the existence of an operator Θ
swith the above-mentioned properties. A key
point of the proof is the truncated observability inequality established in Sec-
tion 3. Subsection 2.3 deals with the construction of an exponentially stabilising
feedback law. Its properties mentioned in the theorem are established in Sub-
section 2.4. In what follows, the domain Ω and its closed subset ω are assumed
to be fixed, and we do not follow the dependence of other quantities on them.
2.2 Construction of a stabilising control
Proposition 2.4. Let the hypotheses of Theorem 2.3 hold and let T > 0 be sufficiently large. Then, for a sufficiently small σ > 0, there is a constant C and an integer m ≥ 1, depending only R and r, such that, for any s ≥ 0, one can construct a continuous linear operator Θ
s: H → L
2(J
T(s), H
σ) satisfying the following properties.
Boundedness. The norm of Θ
sis bounded by C for any s ≥ 0, and its image is contained in L
2(J
T(s), F
m).
Squeezing. Let [v
0, v
1] ∈ H and η = Θ
s[v
0, v
1]. Then the solution of (1.7), (1.8), (2.2) satisfies inequality (2.4).
An immediate consequence of this proposition is the following result on the existence of a stabilising control. For β > 0 and a Banach space X , we denote by L
2β(R
+, X) the space of locally square-integrable functions f : R
+→ X such that
kf k
L2β
:= sup
t≥0
Z
t+1 te
βskf (s)k
2Xds < ∞.
Corollary 2.5. Under the hypotheses of Theorem 2.3, there is β > 0 and a continuous linear operator Θ : H → L
2β(R
+, F
m), where F
mis endowed with the norm of H
σwith some σ > 0, such that the solution of problem (1.7), (1.8), (2.2) with η = Θ[v
0, v
1] and s = 0 satisfies inequality (2.3) with s = 0.
Proof. Let us define a control η : R
+→ F
mby relations (2.5). It follows from (2.4) that
E
v(kT ) ≤ e
−βT kE
v(0), k ≥ 0. (2.6) Since the norms of Θ
sare bounded uniformly in s ≥ 0, it follows from (2.5) and (2.6) that
η|
JT(kt)L2(JT(kT),Fm)
≤ C kΦ
v(kT )k
H≤ Ce
−βT k/2[v
0, v
1]
H. (2.7) This inequality shows that η ∈ L
2β(R
+, F
m) and the operator [v
0, v
1] 7→ η is continuous from H to L
2β(R
+, F
m). Furthermore, the continuity of the resolving operator for problem (1.7), (1.8) and inequalities (2.6) and (2.7) imply that
sup
t∈JT(kT)
kΦ
v(t)k
H≤ C
Φ
v(kT )
H+
η|
JT(kt)L2(JT(kT),Fm)
≤ Ce
−βT k/2[v
0, v
1]
H.
This immediately implies the required estimate (2.3) with s = 0.
Proof of Proposition 2.4. We first describe the scheme of the proof. Define an energy-type functional for a trajectory v(t, x) by the relation
E
v(t) = Z
Ω
| v| ˙
2+ |∇v|
2+ αv v ˙
dx.
For small α > 0, this quantity is equivalent to E
v(t):
C
−1E
v(t) ≤ E
v(t) ≤ CE
v(t), t ∈ R
+. (2.8) Let z be the solution of problem (1.7), (1.8), (2.2) with b ≡ η ≡ 0. Taking the scalar product in L
2of the equation for z with 2 ˙ z + αz, we can find δ > 0 such that
E
z(t) ≤ CE
z(t) ≤ Ce
−δ(t−s)E
z(s) ≤ C
2e
−δ(t−s)E
z(s). (2.9) In particular, if T > 0 is sufficiently large, then
kΦ
z(s + T )k
H≤ 1 4
[v
0, v
1]
H. (2.10)
We seek a solution in the form v = z + w. Then w must be a solution of the control problem
¨
w + γ w ˙ − ∆w + b(t, x)w = η(t, x) − b(t, x)z, x ∈ Ω, (2.11) w
∂Ω= 0, (2.12)
w(s, x) = 0, w(s, x) = 0. ˙ (2.13)
Given an integer N ≥ 1 and a constant ε > 0, we shall construct a control η such that the corresponding solution w satisfies the inequalities
kΦ
w(s + T )k
Hσ≤ M
5[v
0, v
1]
H, kP
NΦ
w(s + T )k ≤ ε
[v
0, v
1]
H, (2.14) where P
Nstands for the orthogonal projection in L
2(Ω) to the vector span of the first N eigenfunctions of the Dirichlet Laplacian, and σ > 0 and M
5are constants not depending on N and ε. For an appropriate choice of N and ε, these two inequalities imply that
kΦ
w(s + T )k
H≤ 1 4
[v
0, v
1]
H. (2.15)
Combining this with (2.10), we see that kΦ
v(s + T )k
H≤ 1
2
[v
0, v
1]
H. This inequality is equivalent to (2.4) with β = T
−1log 2.
We now turn to the accurate proof. The derivation of inequality (2.10) is classical (e.g., see Section 6 in [BV92, Chapter 2]), and we shall confine ourselves to the construction of w. To simplify notation, we shall assume that s = 0; the case s > 0 can be treated by a literal repetition of the argument used for s = 0.
Step 1. We seek η in the form
η(t, x) = χ(x) P
mζ(t, ·)
, (2.16)
where ζ ∈ L
2(J
T× Ω) is an unknown function and m ≥ 1 is an integer that will be chosen later. Let us define the space
X
T= C(J
T, H
01) ∩ C
1(J
T, L
2) ∩ W
2,2(J
T, H
−1)
and consider the following minimisation problem:
Problem 2.6. Given initial data [v
0, v
1] ∈ H and (small) positive numbers δ and σ, minimise the functional
J (w, ζ) = 1 2
Z
T 0kζ(t, ·)k
2σdt + 1
δ k∇ P
Nw(T )k
2+ k P
Nw(T ˙ )k
2in the class of functions (w, ζ) ∈ X
T× L
2(J
T, H
σ) satisfying Eqs. (2.11), (2.13) with s = 0 and η given by (2.16).
This is a linear-quadratic optimisation problem, and it is straightforward to prove the existence and uniqueness of an optimal solution, which will be denoted by (w, ζ). The mapping z 7→ (w, ζ ) is linear, and therefore so is the mapping [v
0, v
1] 7→ η. Let us derive some estimates for the norms of w and ζ. To this end, we write the optimality conditions:
¨
q − γ q ˙ − ∆q + b(t, x)q = 0, (2.17)
(−∆)
σζ = P
m(χq), (2.18)
q(T ) = − 2
δ P
Nw(T ˙ ), q(T ˙ ) = − 2
δ P
N∆w(T ) + γ w(T ˙ )
, (2.19)
where q ∈ L
2(J
T, H
01) is a Lagrange multiplier. Note that, in view of (2.17) and the uniqueness of a solution for the linear wave equation, the function q must belong to X
T, so that relations (2.19) make sense. Let us take the scalar product in L
2(J
T× Ω) of Eqs. (2.11) and (2.17) with the functions q and w, respectively, and take the difference of the resulting equations. After some simple transformations, for σ ∈ (0,
12) we obtain
( ˙ w(T ) + γw(T ), q(T )) − (w(T ), q(T ˙ )) + Z
T0
(bz, q) dt − Z
T0
ζ, P
m(χq) dt = 0.
Using (2.18) and (2.19), we obtain 2
δ k P
Nw(T )k
21+ k P
Nw(T ˙ )k
2+
Z
T 0P
m(χq)
2−σ
dt = Z
T0
(bz, q) dt. (2.20) There is no loss of generality in assuming that T is so large that inequal- ity (3.2) holds and, hence, the truncated observability inequality (3.11) is true for small σ > 0. Combining this with (1.16), (1.13), and (2.9) we derive
|(bz, q)| ≤ kbzk
σkqk
−σ≤ Ckzk
1kΦ
q(0)k
H−σ≤ CkΦ
z(0)k
HP
m(χq)
L2(JT,H−σ). The Cauchy–Schwarz inequality now implies that
Z
T 0(bz, q) dt ≤ 1
2 Z
T0
P
m(χq)
2−σdt + C kv
0k
21+ kv
1k
2.
Substituting this into (2.20), we obtain 1
δ k P
Nw(T )k
21+ k P
Nw(T ˙ )k
2+
Z
T 0P
m(χq)
2−σ
dt ≤ C
[v
0, v
1]
2H
, (2.21) where m = m(N ) ≥ 1 is an integer and C is a constant not depending on δ and N. Taking N ≫ 1 and δ ≪ 1, we obtain the second inequality in (2.14) with s = 0.
Step 2. Let un prove the boundedness of the operator Θ
0: [v
0, v
1] 7→ η from H to L
2(J
T, H
σ) and inequality (2.15) with s = 0. This will complete the proof of Proposition 2.4.
It follows from (2.18) and (2.21) that kζk
L2(JT,Hσ)≤ C
[v
0, v
1]
H.
Since the projection P
mand multiplication by χ are bounded operators in H
σ, the above inequality combined with relation (2.16) shows that Θ
0is bounded.
To prove (2.15), we write
kΦ
w(T )k
H≤ k P
NΦ
w(T )k
H+ k(I − P
N)Φ
w(T )k
H≤ ε
[v
0, v
1]
H+ δ
N(σ)kΦ
w(T )k
Hσ,
where δ
N(σ) → 0 as N → ∞. It follows that (2.15) will be established if we prove the first inequality in (2.14).
Duhamel representation for solutions of (2.11)–(2.13) and inequality (1.12) with s = σ imply that
kΦ
w(T )k
Hσ≤ C
η − bz − bw
L1(JT,Hσ).
Combining this with (1.16) and using condition (2.1) and the boundedness of z and w in C(J
T, H
01), we arrive at the required inequality.
2.3 Dynamic programming principle and feedback law
Once the existence of a stabilising control is established, an exponentially stabil- ising feedback law can be constructed using a standard approach based on the dynamic programming principle. Since the corresponding argument was carried out in detail for the more complicated case of the Navier–Stokes system (see Section 3 in [BRS11]), we shall omit some of the proofs. Let us consider the following optimisation problem depending on the parameter s ≥ 0.
Problem 2.7. Given [v
0, v
1] ∈ H and β > 0, minimise the functional I
s(v, ζ) = 1
2 Z
∞s
e
βtk∇v(t)k
2+ k v(t)k ˙
2+ kζ(t)k
2dt in the class of functions (v, ζ) such that
v ∈ C(R
s, H
01) ∩ C
1(R
s, L
2) ∩ W
loc2,2(R
s, H
−1), ζ ∈ L
2loc(R
s, L
2),
and Eqs. (1.7) and (2.2) hold with η given by (2.16).
This is a linear-quadratic optimisation problem, and in view of Corollary 2.5, there is at least one admissible pair (v, ζ) for which I
s(v, ζ) < ∞. It follows that there is a unique optimal solution (v
s, ζ
s) for Problem 2.7, and the corresponding optimal cost can be written as
I
s(v
s, ζ
s) = 1
2 Q
s[v
0, v
1], [v
0, v
1]
H
,
where Q
s: H → H is a bounded positive operator in the Hilbert space H.
Moreover, repeating the argument used in the proof of Lemma 3.8 in [BRS11], one can prove that Q
scontinuously depends on s in the weak operator topology, and its norm satisfies the inequality
kQ
sk
L(H)≤ C e
βs, s ≥ 0. (2.22) We now consider the following problem depending on the parameter s > 0.
Problem 2.8. Given [v
0, v
1] ∈ H and β > 0, minimise the functional K
s(v, ζ) = 1
2 Z
s0
e
βtk∇v(t)k
2+ k v(t)k ˙
2+ kζ(t)k
2dt + 1
2 Q
s[v
0, v
1], [v
0, v
1]
H
in the class of functions (v, ζ) such that
v ∈ C(J
s, H
01) ∩ C
1(J
s, L
2) ∩ W
2,2(J
s, H
−1), ζ ∈ L
2(J
s, L
2), and Eqs. (1.7) and (1.9) hold with η given by (2.16).
This is a linear-quadratic optimisation problem, which has a unique solu- tion (˜ v
s, ζ ˜
s). The following lemma establishes a link between Problems 2.7 and 2.8. Its proof repeats the argument used in [BRS11] (see Lemma 3.10) and is omitted.
Lemma 2.9. Let (v, ζ) = (v
0, ζ
0) be the unique solution of Problem 2.7 with s = 0. Then the restriction of (v, ζ) to the interval J
scoincides with (˜ v
s, ζ ˜
s) and the restriction of (v, ζ) to the half-line R
scoincides with (v
s, ζ
s) corresponding to the initial data [v(s), v(s)]. ˙
The optimality conditions for the restriction of (v, ζ) to J
simply, in partic- ular, that
¨
q
s− γ q ˙
s− ∆q
s+ b(t, x)q
s= e
βt∆v + β v ˙ + ¨ v
, (2.23)
ζ(t) = e
−βtP
m(χq
s(t)), (2.24)
q
s(s) = −Q
1sΦ
v(s), (2.25)
where q
s∈ L
2(J
s, H
01) is a Lagrange multiplier and Q
1s: H → L
2is a contin-
uous operator defined by the relation Q
sV = [Q
0sV, Q
1sV ] for V ∈ H. Since
the right-hand side of (2.23) belongs to L
2(J
s, H
−1), it follows from Propo-
sition 1.3 with σ = 0 that the function q
smust belong to C(J
s, L
2), so that
relation (2.25) makes sense, and ζ is a continuous function of time with range in F
m. Combining (2.24) and (2.25), we see that
ζ(s) = −e
−βsP
mχQ
1sΦ
v(s) .
Recalling that s > 0 is arbitrary and using (2.16), we conclude that the unique optimal solution (v, ζ) of Problem 2.7 with s = 0 satisfies Eq. (1.7) with η(t, x) = K
b(t)Φ
v(t), where the linear operator K
b(t) : H → F
mis given by
K
b(t)V = −e
−βtχ P
mχQ
1tV
, t ≥ 0. (2.26)
In the next section, we shall show that this operator satisfies all the properties mentioned in Theorem 2.3.
2.4 Conclusion of the proof of Theorem 2.3
The continuity of the function t 7→ K
b(t) in the weak operator topology follows from a similar property for Q
t, and the uniform boundedness of its norm is an immediate consequence of (2.22). To establish (2.3), we first consider the case s = 0. Let us define w(t, x) = e
βt/2v(t, x). Then there is C > 1 such that
C
−1E
w(t) ≤ e
βtE
v(t) ≤ C E
w(t), t ≥ 0. (2.27) Furthermore, the function w must satisfy the equation
¨
w + γ w ˙ − ∆w = g(t, x), (2.28) where we set
g(t, x) = e
βt/2K
b(t)Φ
v(t) +
β42+
γβ2− b
v + β v ˙ .
Note that kg(t)k
2≤ C e
βtE
v(t), and since v is the optimal solution of Prob- lem 2.7 with s = 0, we have
Z
∞ 0kg(t)k
2dt ≤ C Q
0[v
0, v
1], [v
0, v
1]
H
≤ C E
v(0). (2.29) Taking the scalar product in L
2of Eq. (2.28) with 2 ˙ w + αw, carrying out some standard transformations, and using the Gronwall inequality, we derive
E
w(t) ≤ Ce
−δtE
w(0) + C Z
t0
e
−δ(t−θ)kg(θ, ·)k
2dθ.
Combining this with (2.27) and (2.29), we arrive at the required inequality (2.3) with s = 0.
To prove (2.3) with an arbitrary s = θ > 0, we repeat the above argu-
ment with the initial point moved to θ. Namely, considering an analogue of
Problem 2.8 on the half-line R
θ, one can prove by the same argument as above that
ζ
θ(t) = −e
−βtP
mχQ
1tΦ
θv(t)
, t > θ.
It follows that if v is the solution of (1.7), (2.2) with η(t) = K
b(t)Φ
v(t) and s = θ, then
Q
θ[v
0, v
1], [v
0, v
1]
H
= 1 2
Z
∞ θe
βtE
v(t) + e
−βtk P
mχQ
1tΦ
v(t) k
2dt
≤ C e
βθE
v(θ).
We can now establish (2.3) by literal repetition of the argument used above for problem on the half-line R
+.
It remains to establish the uniqueness of solution. Let v(t, x) be a func- tion that belongs to the space C(R
+, H
01) ∩ C
1(R
+, L
2) and satisfies Eqs. (1.7) and (2.2) with η(t) = K
b(t)Φ
v(t) and v
0= v
1≡ 0. Since η ∈ L
1loc(R
+, L
2), inequality (1.13) and the boundedness of the operator K
b(t) imply that
kΦ
v(t)k
H≤ C Z
t0
kΦ
v(s)k
Hds.
By the Gronwall inequality, we conclude that v ≡ 0. The proof of Theorem 2.3 is complete.
3 Observability inequalities
This section is devoted to the proof of a truncated observability inequality used in Section 2.2. Namely, let us consider the homogeneous equation
¨
v − γ v ˙ − ∆v + b(t, x)v = 0, x ∈ Ω, (3.1) supplemented with the Dirichlet boundary condition (1.8). We first establish a “full” observability inequality for solutions of low regularity and then use a compactness argument to derive the required result.
3.1 Observability of low-regularity solutions
Theorem 3.1. Let the hypotheses of Theorem 2.3 be fulfilled, let χ ∈ C
0∞(R
3) be such that supp χ ∩ Ω ⊂ ω and χ(x) = 1 for x ∈ Ω
δ/2(x
0), and let
T > 2 sup
x∈Ω
|x − x
0|, (3.2)
where x
0∈ R
3is the point entering Condition 2.2. Then there are positive constants σ
0(r) and M
6= M
6(R, r, T, γ, χ) such that, for any initial data [v
0, v
1] ∈ H
−σwith 0 ≤ σ ≤ σ
0(r) the solution v(t, x) of problem (3.1), (1.8), (1.9) satisfies the inequality
kv
0k
2−σ+ kv
1k
2−σ−1≤ M
6Z
T 0kχv(t)k
2−σdt. (3.3)
Proof. The proof is divided into three steps: we first establish a unique contin- uation property for low-regularity solutions; we then use the Bardos–Lebeau–
Rauch observability inequality to establish a high-frequency observability; and, finally, these two results are combined to prove the required observability in- equality. Without loss of generality, we shall assume that γ = 0; the general case can easily be treated by the change of variable v(t) = e
γt/2w(t).
Step 1. Unique continuation property. Let σ ∈ (0, 1) be so small that the initial-boundary value problem for (3.1) is well posed in H
−σand the conclusion of Lemma 1.6 is true. We claim that if a solution v(t, x) of problem (3.1), (1.8), (1.9) with initial data [v
0, v
1] ∈ H
−σis such that χv = 0 on J
T× Ω, then v ≡ 0.
Indeed, let ϕ ∈ C
0∞(R) be such that ϕ = 1 around 0. For h ∈ (0, 1], let v
hbe the solution of (3.1), (1.8) with the initial condition
v
h t=0= ϕ(−h
2∆)v
0, ∂
tv
h t=0= ϕ(−h
2∆)v
1. We set u
h= v
h− ϕ(−h
2∆)v. Then
∂
t2u
h− ∆u
h+ b(t, x)u
h= [ϕ(−h
2∆), b]v, [u
h, ∂
tu
h]
t=0= 0. (3.4) By Proposition 1.3 (with σ = 0) and Lemma 1.6, for all t ∈ J
Twe have
kΦ
uh(t)k
H0≤ M
2Z
T 0ϕ(−h
2∆), b(s) v(s)
−1ds
≤ M
2M
4Z
T 0h
σkv(s)k
−σds ≤ C h
σk[v
0, v
1]k
H−σ. (3.5) Suppose we have shown that
kΦ
vh(t)k
H0≤ C h
σ+ h
1−σ[v
0, v
1]
H−σ. (3.6)
Then these two inequalities imply that
h→0
lim kϕ(−h
2∆)v
0k + kϕ(−h
2∆)v
1k
−1= 0, whence we conclude that v
0= v
1= 0 and, hence, v ≡ 0.
To prove (3.6), recall that, by [DZZ08], the observability inequality (3.3) is true with σ = 0. Combining this with (1.13), we derive
kΦ
vh(t)k
2H0≤ M
22kΦ
vh(0)k
2H0≤ M
22M
6Z
T 0kχv
h(s)k
2ds
≤ C Z
T0
χu
h(s)
2ds + C Z
T0
χ ϕ(−h
2∆) v(s)
2ds. (3.7) By (3.5), the first term on the right-hand side does not exceed Ch
2σk[v
0, v
1]k
2H−σ. To estimate the second term, we write
χ ϕ(−h
2∆) v = ϕ(−h
2∆)χv
| {z }
=0